Currently, a range of communication channels (including the Internet and wireless channels) can be modeled as erasure channels. It is well known that the transmission of multiple mutually correlated but distinct versions of a given signal enhance the probability of the high fidelity reconstruction of the given signal. The underlying concept of such techniques is that the erasure channels have a more or less independent probability of failure, thus ensuring that the probability of the correct reception of at least a few channels is high even when the correct reception of a single channel is low.
The transmission of several exact replicas of a given signal over a channel may result in an explosion of bandwidth required to transmit the signal over the channel. As illustrated in FIG. 1, a solution to this problem involves transmitting several correlated but not exact replicas of the signal over the channel. When such a technique is employed the reconstruction fidelity of the signal improves with the number of copies of the signal that are correctly received. There are numerous publications relating to the above-mentioned signal transmission problem. Related inventions disclosed in literature include U.S. Pat. No. 6,330,370 issued to Goyal, et al. for the “Multiple description transform coding of images by the use of optimal transforms of arbitrary dimension” and U.S. Pat. No. 6,542,554 issued to Jafarkhani, et al. for “Multiple description trellis coded quantization”.
Although such techniques have been shown to achieve good performance in practice, their extension to the case of predictively encoded signals, i.e., signal coding with memory, is not well understood. If the source signal sequence consists of predictively encoded symbols, as in the case of video coding, the additional problem of predictive mismatch is encountered.
Predictive mismatch refers to the scenario in which the reconstruction of the predictor symbol at the decoder is different from the predictor symbol used at the encoder. This causes the reconstruction of the predicted symbol to be erroneous and this effect propagates through the sequence. In the context of transmission of multiple correlated signals over independent channels, since the number of channel failures in transmitting the predictor symbol is unknown at the encoder, the decoder reconstruction of the predictor cannot be accurately reproduced at the encoder leading to a mismatch. Thus, the key issue to be solved in the case of predictively encoded symbols is circumventing the problem of predictive mismatch.
Layered coding is a coding technique that uses multiple channels for the transmission of predictively coded information in the presence of channel erasures. However, layered coding requires that one particular channel, which is typically referred to as the ‘base channel’, has to be received with fidelity in order for a scheme to succeed. The present invention does not rely on such restrictions and therefore possesses greater general applicability. The resultant problem when two correlated representations are employed is illustrated in FIG. 2.
The residual-of-residuals (RoR) technique addresses this problem by sending a correction signal that partially removes predictive mismatch. However, the increase in rate incurred in transmitting the correction signal is quite high. U.S. Pat. No. 6,556,624 to Orchard, et al. for a “Method and apparatus for accomplishing multiple description coding for video” discloses an invention that utilizes a similar technique for video transmission. It must be noted that this technique, in direct contrast to the present invention, requires the transmission of at least one additional correction signal to accomplish its objectives.
The invention disclosed in U.S. Pat. No. 6,611,530 to Apostolopoulos for “Video communication using multiple streams” discloses a method for transmitting predictively encoded video using multiple streams, such that errors in one or more streams does not destroy the entire video stream but only reduces the play back frame rate. Conversly, the utilization of the present invention does not lead to a reduction in the frame rate in the event of the loss of one or more channels.
The present invention eliminates predictive mismatch while avoiding transmission of a correction signal.